Strategy-Proofness makes the Difference:
Deferred-Acceptance with Responsive Priorities
∗Lars Ehlers† Bettina Klaus‡ September 2012
Abstract
In college admissions and student placements at public schools, the admission decision can be thought of as assigning indivisible objects with capacity constraints to a set of students such that each student receives at most one object and monetary compensations are not allowed. In these important market design problems, the agent-proposing deferred-acceptance (DA-)mechanism with responsive strict priorities performs well and economists have successfully implemented DA-mechanisms or slight variants thereof. We show that almost all real-life mechanisms used in such environments—including the large classes of priority mechanisms and linear programming mechanisms—satisfy a set of simple and intuitive properties. Once we add strategy-proofness to these properties, DA-mechanisms are the only ones surviving. In market design problems that are based on weak priorities (like school choice), generally multiple tie-breaking (MTB) procedures are used and then a mechanism is implemented with the obtained strict priorities. By adding stability with respect to the weak priorities, we establish the first normative foundation for MTB-DA-mechanisms that are used in NYC.
JEL Classification: D63, D70
Keywords: deferred-acceptance mechanism, indivisible objects allocation, multiple tie-breaking, school choice, strategy-proofness.
1
Introduction
In centralized entry-level medical markets students search for their first positions at hospitals. Each year first the finishing students submit a ranking of the available hospital positions and the
∗
Lars Ehlers acknowledges the hospitality of Toulouse School of Economics (INRA-LERNA), where an early version of this paper was written, and financial support from the SSHRC (Canada) and the FQRSC (Qu´ebec). Bettina Klaus acknowledges the hospitality of Harvard Business School, where part of this paper was written, and financial support from the Netherlands Organisation for Scientific Research (NWO) under grant VIDI-452-06-013. This article is a substantially revised version of Ehlers and Klaus (2009).
†
Corresponding author : D´epartement de Sciences ´Economiques and CIREQ, Universit´e de Montr´eal, Montr´eal, Qu´ebec H3C 3J7, Canada; e-mail: lars.ehlers@umontreal.ca.
‡
Faculty of Business and Economics, University of Lausanne, Internef 538, CH-1015, Lausanne, Switzerland; e-mail: bettina.klaus@unil.ch.
hospitals submit a ranking of the finishing students (and hospitals’ preferences over sets of students are supposed to be “responsive” with respect to these ranking). Then, the centralized clearinghouse matches via a mechanism students and hospitals on the basis of the submitted lists. A matching is stable if (i) every student prefers his match to being unmatched, (ii) every hospital prefers each matched student to having the position unfilled, and (iii) there is no student-hospital pair that blocks the matching in the sense that the student prefers the hospital to his match and the hospital prefers the student to one of its matched students or having the position unfilled.
The American markets (Roth, 1984a) and the British markets in Edinburgh and Cardiff based the matching on a stable mechanism called hospital-proposing deferred-acceptance (DA-)mechanism (Gale and Shapley, 1962). Although the DA-mechanism is the predominant mechanism in the two-sided matching market literature (Roth, 2008), British markets have introduced and operated unstable mechanisms (Roth, 1991). The markets in Edinburgh (initially), Newcastle, and Birm-ingham have adopted priority mechanisms in 1966 and 1967, and the markets in Cambridge and the London Hospital Medical College have adopted linear programming mechanisms in the 1970’s. The literature on two-sided matching markets has argued that in such markets stable mechanisms are more successful than others because unstable mechanisms may produce matchings which are blocked and unsustainable (e.g., Roth, 1991). This is especially due to the fact that both sides of the market have to be considered as active agents who, after the announced outcome of the centralized clearinghouse, can resign from positions (students) or fire students (hospitals) in order to engage in new matches.
In all these markets preferences are private information and agents need to reveal them. Ideally we would like to base the outcome on the true preferences and construct an incentive compatible mechanism. Unfortunately there does not exist any stable mechanism that, for all students and for all hospitals, makes truth-telling a weakly dominant strategy (Roth, 1982). However, the student-proposing DA-mechanism is strategy-proof for students (Dubins and Freedman, 1981; Roth, 1982), but there does not exist any stable mechanism that is strategy-proof for hospitals with several positions (Roth, 1984b). This is one of the main reasons why in the mid 1990’s the American mar-kets changed the mechanism from the hospital-proposing DA-mechanism to the student-proposing DA-mechanism.
Over the past couple of years, the two-sided matching problem has been adapted to address the assignment of students to public schools. In a so-called school choice problem, schools correspond to hospitals and students seek a slot at a school (Abdulkadiro˘glu and S¨onmez, 2003). The difference to a two-sided matching market problem is that a school’s “preference relation” represents priorities of students to be admitted at that school. These priorities over individual agents are fixed (via exogenous criteria or objective tests) and cannot be changed. Furthermore, schools are not strategic agents. Now if priorities are strict (and priorities over sets of students are responsive with respect to the priorities over individual agents), then the DA-mechanism can be applied in a straightforward
manner. The same is true for priority mechanisms and linear programming mechanisms. Indeed, the famous Boston mechanism for school choice is equivalent to the Edinburgh 1967 priority mechanism (Ergin and S¨onmez, 2006) and priority mechanisms are nowadays still in use in school choice.1
In school choice, schools are not strategic agents and the motivation for stability needs to be slightly reconsidered (a school simply cannot deny enrollment to a student who has been as-signed to it by the central authorities in order to rematch with a preferred student). The intuitive reinterpretation of stability for school choice problems is that stability property (i) still represents students’ individual rationality, stability property (ii) corresponds to the non-wastefulness of school seats, and stability property (iii) captures the fairness property of no justified envy2 (Balinski and S¨onmez, 1999). Next, several contributions underline the policy importance of strategy-proofness for school choice mechanisms. The argument there is not simply that truthful revelation of pref-erences is desirable, but that it is important to level the playing field for students from different backgrounds. Abdulkadiro˘glu, Pathak, Roth, and S¨onmez (2006) and Pathak and S¨onmez (2008) show that under the Boston mechanism sophisticated agents were much better off than unsophis-ticated agents: na¨ıvely truth-telling students (or parents) tend to be the worst off students under the Boston mechanism, which is not strategy-proof. This is due to the fact that with a manip-ulable mechanisms it may be difficult to figure out the best (non-truthful) strategy whereas with a strategy-proof mechanisms each agent’s simple strategy is to report preferences truthfully. In this paper we formulate some simple and intuitive properties that are satisfied by many mecha-nisms (using some arbitrary strict priorities) including priority mechamecha-nisms, linear programming mechanisms and responsive DA-mechanisms. However, we show that once strategy-proofness is added, only the student-proposing DA-mechanisms (with responsive priorities) survive among all mechanisms satisfying our set of simple and intuitive properties (Theorem 1). This provides a rationale for using the student-proposing responsive DA-mechanism if truthful revelation of prefer-ences is important. Furthermore, the feature of responsiveness-to-individual-priorities makes this mechanism easily applicable in practice (Roth, 2008).3
For some school choice problems priorities may be coarse, e.g., in Boston priorities are divided into four classes: (i) siblings and walk zone, (ii) siblings, (iii) walk zone, and (iv) the rest. In extreme cases, no priorities are given and all students have equal rights for receiving any object. When priorities are weak, in order to apply a (student-proposing) DA-mechanism, one needs to resolve the ties in some way. Abdulkadiro˘glu, Pathak, and Roth (2009) propose to use a fixed
1In Boston this mechanism has now been changed to the student-proposing DA-mechanism, but many cities
worldwide still use the so-called Boston (or direct acceptance) mechanism.
2A matching eliminates justified envy if whenever a student prefers another student’s match to his own, he has a
lower priority than the other student at that school.
3
The search for “good” mechanisms is the subject of many recent contributions, but most of them deal with the house allocation model where exactly one object of each type is available (e.g., Ehlers, 2002; Ehlers and Klaus, 2006, 2007; Kesten, 2009; P´apai, 2000). In most papers that study the allocation of indivisible objects with capacity constraints, externally prescribed priorities are also specified; the corresponding class of problems is usually referred to as “school choice problems” or “student placement problems” (see S¨onmez and ¨Unver, 2011, for a recent survey).
multiple tie-breaking procedure (MTB) (which is independent of students’ preferences) and then apply the DA-mechanism. Erdil and Ergin (2008) propose to use different tie-breaking procedures for different preference profiles. Both these contributions regard stability (with respect to the weak priorities) as an important requirement. We show that in Theorem 1, we can replace individual rationality and weak non-wastefulness with stability with respect to weak priorities, and by doing so we establish the first normative foundation of responsive MTB-DA-mechanisms as used in NYC (Theorem 2).
The paper is organized as follows. In Section 2 we introduce a model of allocating indivisi-ble objects with capacity constraints and exogenously given priorities. In Section 3 we introduce tie-breaking procedures for exogenously given priorities, priority mechanisms, linear programming mechanisms, and responsive DA-mechanisms. In Section 4 we state simple and intuitive properties for mechanisms. In Section 5 we show that all described real-life mechanisms satisfy our simple and intuitive properties. We then show that only (agent-proposing) responsive DA-mechanisms satisfy these simple and intuitive properties and strategy-proofness. For exogenously given pri-orities, in Section 6, we then obtain a characterization of responsive MTB-DA mechanisms when imposing stability. The Appendix contains all our proofs and in addition a second characterization of responsive DA-mechanisms with weak consistency.4
2
Allocation with Priorities and Variable Resources
Let N = {1, . . . , n} denote a finite set of agents with n ≥ 2. Let O denote a set of potential (real) object types or types for short. We assume that O contains at least two elements and that O is finite.5 Not receiving any real object is called “receiving the null object.” Let ∅ represent the null object.
Each agent i ∈ N is equipped with a preference relation Riover all types O∪{∅}. The preference
relation Ri is strict, i.e., Ri is a linear order over O ∪ {∅}. Given x, y ∈ O ∪ {∅}, x Piy means that
agent i strictly prefers x to y (and x 6= y) and x Riy means that agent i weakly prefers x to y
(and x Piy or x = y). Let R denote the set of all preferences over O ∪ {∅}, and RN the set of all
(preference) profiles R = (Ri)i∈N such that for all i ∈ N , Ri ∈ R.
Given R ∈ RN and M ⊆ N , let RM denote the profile (Ri)i∈M. It is the restriction of R to the
set of agents M . We also use the notation R−M = RN \M and R−i= RN \{i}.
Given O0 ⊆ O ∪ {∅}, let Ri|O0 denote the restriction of Ri to O0 and R|O0 = (Ri|O0)i∈N. Given
i ∈ N and Ri ∈ R, type x ∈ O is acceptable under Ri if x Pi∅. Let A(Ri) = {x ∈ O : x Pi∅} denote
the set of acceptable types under Ri.
4
Our Supplementary Appendix also discusses Kojima and Manea (2010) in relation to our results and provides the independence of the properties used in our characterizations.
5
For each type x ∈ O, at most qx ∈ N copies are available in any economy with 1 ≤ qx ≤ |N |.6
Let qx, 0 ≤ qx ≤ qx, denote the number of available objects or the capacity of type x. Let
q = (qx)x∈O denote a capacity vector and Q = ×x∈O{0, 1, . . . , qx} denote the set of all capacity
vectors. The null object is always available without scarcity and therefore we set q∅ = ∞.
Given type x, letbx denote a (weak) priority ordering of type x on N . Here i bxj means that agent i has higher priority than agent j for obtaining an object of type x and i∼bxj means that agents i and j have the same priority for obtaining an object of type x. For each type x, the priority ordering bx is exogenously given. The priority ordering bx is strict if for any distinct i, j ∈ N , we have either i bx j or j bx i. Let = (b bx)x∈O denote a priority structure (e.g., determined
via test/exam scores or other objective suitability criteria). Priority structure is strict if for allb x ∈ O, priority orderingbx is strict. For a strict priority structure we sometimes writeb insteadb of . In the sequel, we denote an exogenously given priority structure byb and an arbitraryb strict priority structure by . In some environments no priorities are given (no tests/exams are conducted). Then all agents have equal priority for any type. We denote the “no or equal priorities” priority structure by∼, i.e., for all types x and for all i, j ∈ N , ib ∼bxj.
An (allocation) problem (with capacity constraints and priorities) consists of a preference profile R ∈ RN, a capacity vector q ∈ Q, and a priority structure . Sinceb is exogenously given andb remains fixed throughout, we often denote a problem by (R, q) and the set of all problems by RN × Q. Given a capacity vector q, let O
+(q) = {x ∈ O : qx > 0} denote the set of available
real types under q. The set of available types is O+(q) ∪ {∅} and includes the null object, which is
available at any problem.
Each agent i is to be allocated exactly one object in O ∪ {∅} taking capacity constraints into account. An allocation for (capacity vector) q is a list a = (ai)i∈N such that for all i ∈ N ,
ai ∈ O ∪ {∅}, and any real type x ∈ O is not assigned more than qx times, i.e., for all x ∈ O,
|{i ∈ N : ai = x}| ≤ qx. Note that ∅, the null object, can be assigned to any number of agents
and that not all real objects have to be assigned. Let Aq denote the set of all allocations for q and
A =S
q∈QAq set of all allocations. Given a problem (R, q) and a ∈ Aq, allocation a is individually
rational if for all i ∈ N , aiRi∅.
An mechanism is a function ϕ : RN × Q → A such that for all problems (R, q) ∈ RN × Q,
ϕ(R, q) ∈ Aq. Given i ∈ N , we call ϕi(R, q) the allotment of agent i at ϕ(R, q).
3
Multiple Tie-Breaking and Real Life Mechanisms
For the application of the real-life mechanisms we introduce in this section, we first have to resolve or break ties. In other words, these mechanisms require the transformation of a weak priority
6By introducing “global upper bounds” via q
x (x ∈ O) we can, for instance, specify the house allocation model
structure into a strict priority structure . In applications this is typically done by breakingb the ties for any type. A strict priority structure is a multiple tie-breaking (MTB) of the (weak) priority structure if for any type x and any i, j ∈ N : if ib bxj, then i xj. Note that then any tie in is broken and respects all strict priorities ofb . Any MTB ofb is strict and we oftenb use the terminology “for any MTB of .”b
3.1 Priority Mechanisms
Below we describe the large class of priority mechanisms. Any priority mechanism, for each prefer-ence profile R and any MTB of, matches sequentially all student-type pairs that have highestb “priority” among all remaining pairs. First, the priority mechanism checks whether there are (1,1)-matches (i.e., an agent-type pair who ranks each other first). If there are any (1,1)-(1,1)-matches, then they are realized. Second, the priority mechanism checks whether there are matches that have the second highest priority (i.e., (1,2)- or (2,1)-matches) and matches any such pair. And so on. Of course, a match is only realized if the agent prefers his matched object to the null object. Be-cause priority mechanisms differ in how they exactly prioritize matches, we have just sketched a class of mechanisms (note the indeterminacy in the above description when mentioning (1,2)- or (2,1)-matches as having the second highest priority). In particular, the Edinburgh 1967 mechanism ordered matches lexicographically according to the types’ priorities, that is, (1,1)-matches have first priority, followed by (2,1), (3,1), and so on. Only when all types’ first choices are exhausted, other matches ((1,2), (2,2), (3,2), . . .) are considered. We will call any mechanism that orders matches lexicographically in types’ priorities or in students’ preferences, a lexicographic priority mechanism. The Boston mechanism for public school choice is a lexicographic priority mechanism (Ergin and S¨onmez, 2006).
Given an agent’s preference relation Ri, we say that agent i’s most preferred real type has rank 1,
his second most preferred real type has rank 2, etc.. For each x ∈ O, we denote by rank(x, Ri) the
rank of x in Ri. Note that we do not rank the null object. Similarly we define rank(i, x). Given
(R, ), we call (i, x) an (a, b)-match if rank(x, Ri) = a and rank(i, x) = b.
A priority function is a one-to-one function g : N × N → N such that (i) for all (a, b), (a0, b) ∈ N × N, if a < a0, then g(a, b) < g(a0, b), and (ii) for all (a, b), (a, b0) ∈ N × N, if b < b0, then g(a, b) < g(a, b0).
The function g associates with each (a, b)-match its priority g(a, b). Hence, in our terminology an (a, b)-match has higher priority than an (a0, b0)-match if g(a, b) < g(a0, b0).
For each problem (R, q), we will only consider the ranks of available real types under q. Thus, the priority function will take as inputs the rank information based on the reduced preference profile R|O+(q)∪{∅}. Then, for each problem (R, q) and any MTB of , the MTB-priority mechanismb
based on g and operates on the restricted problem (R|O+(q)∪{∅}, q) and matches sequentially all
agent-type pairs that have the highest priority according to g subject to individual rationality and the quotas. Let M(g, )(R, q) denote the resulting allocation and let M(g, ) denote the priority
mechanism based on g and .
3.2 Linear Programming Mechanisms
A linear programming mechanism assigns to each possible (a, b)-match a positive weight. These weights are strictly decreasing in both components. A weighting function is a function h : (N × N) ∪ {(0, 0)} → R+ such that
(i) h(0, 0) = 0,
(ii) for all (a, b), (a0, b) ∈ N × N, if a < a0, then h(a, b) > h(a0, b), and (iii) for all (a, b), (a, b0) ∈ N × N, if b < b0, then h(a, b) > h(a, b0).
Recall that rank(x, Ri) denotes the rank of real type x ∈ O in Ri and rank(i, x) the rank of
i ∈ N in x. For all i ∈ N , we now also define rank(∅, Ri) = 0 and rank(i, ∅) = 0. Given a
profile R ∈ R and an individually rational allocation a, the score of a is the sum of the weights of the matched agent-type pairs, i.e.,
s(h, )(a, R) = X
i∈N
h(rank(ai, Ri), rank(i, ai)).
Given a weighting function h, the linear programming mechanism M(h, ) chooses for each problem (R, q) an individually rational allocation for q with maximal score among all individually rational allocations for q. Let arg max s(h, )(·, R) denote the set of all individually rational alloca-tions for q that maximize the weighted score s(h, )among all individually rational allocations for q. We will assume that if there are several allocations maximizing the score under a submitted profile, then the linear programming mechanism breaks ties uniformly, i.e., each of these maximizers is chosen with equal probability. In such a case a linear programming mechanism chooses a lottery over allocations. It is easy to see that this does not create any significant problem and our analysis can be extended to mechanisms choosing a lottery for each profile.
Let M denote a random mechanism choosing for each problem (R, q) a lottery over Aq. Given
an allocation a for q, let Pr{M (R, q) = a} denote the probability that M assigns to allocation a. Similarly, given i ∈ N and y ∈ O ∪ {∅}, let Pr{Mi(R, q) = y} denote the probability that i is
assigned to y under the lottery M (R, q).
For each problem (R, q), we will only consider the ranks of available real types under q. Thus, the weighting function will take as inputs the rank information based on the reduced preference
profile R|O+(q)∪{∅}. Then, for each problem (R, q) and any MTB of, the MTB-linear program-b
ming mechanism based on h and operates on the restricted problem (R|O+(q)∪{∅}, q) and for all
individually rational allocations a for q such that s(h, )(a, R|O+(q)∪∅) = max
a0∈A
qis individually rational
s(h, )(a0, R|O+(q)∪∅), we have
Pr{M(h, )(R, q) = a} = | arg max s(h, )1(·,R|
O+(q)∪∅)| (where |S| denotes the cardinality of set S).
3.3 Stability and Agent-Proposing Responsive DA-Mechanisms
Given the priority structure and a problem (R, q), we can interpret (R,b , q) as a college ad-b missions problem (Gale and Shapley, 1962; Roth and Sotomayor, 1990) where the set of agents N corresponds to the set of students, the set of real types O corresponds to the set of colleges, the capacity vector q describes colleges’ quota, preferences R correspond to students’ preferences over colleges and not receiving any college, and the priority structure is taken to represent col-b leges’ responsive preferences over students. Stability is an important requirement for many real-life matching markets and it turns out to be essential in our context of allocating indivisible objects to agents.
Stability under : Given (R, q) ∈ Rb N× Q, an allocation a ∈ Aq is stable under if there existsb no agent-type pair (i, x) ∈ N × O ∪ {∅} such that x Piai and (s1) |{j ∈ N : aj = x}| < qx or
(s2) there exists k ∈ N such that ak = x and ib xk.
Note that (s1) implies individual rationality because q∅ = ∞. Furthermore, mechanism ϕ is
stable if there exists a priority structure such that for each problem (R, q) ∈ RN × Q, ϕ(R, q) is stable under . When is strict, we suppose that colleges’ priorities over sets of students are responsive with respect to (we discuss responsiveness in detail below in Remark 1).
For any college admissions problem with strict and responsive priorities (R, , q), we denote by DA (R, q) the agent-optimal stable allocation that is obtained by using Gale and Shapley’s (1962) agent-proposing deferred-acceptance algorithm, DA-algorithm for short:7 let (R, q, ) be given. Then,
• at the first step of the DA-algorithm, every agent applies to his favorite acceptable type. For each real type x, the qx applicants who have the highest priority for x (all applicants if there
are fewer than qx) are placed on the waiting list of x, and all others are rejected.
• At the r-th step of the DA-algorithm, those applicants who were rejected at step r − 1 apply to their next best acceptable type. For each real type x, the qx applicants among the new
7
Analogously we may define the type-proposing deferred-acceptance algorithm where each type applies at each step to as many agents as there are objects of this type available.
applicants and those on the waiting list who have the highest priority for x are placed on the updated waiting list of x, and all others are rejected.
The DA-algorithm terminates when every agent is on a waiting list or has applied to all ac-ceptable types. Once the algorithm ends, objects are assigned to the agents on the respective type waiting lists and the null object is assigned to the agents who are not on any waiting list. The resulting allocation is the agent-optimal stable allocation for the problem (R, q, ), denoted by DA (R, q).
Responsive DA-Mechanisms: A mechanism ϕ is a responsive DA-mechanism if there exists a strict priority structure such that for each (R, q) ∈ RN × Q, ϕ(R, q) = DA (R, q).
In real life we often take a MTB of before applying the DA-mechanism DAb . We will call such a mechanism a responsive MTB-DA-mechanism.
Note that the above condition of stability under takes only blocking by individual agents and by agent-type pairs into account. This is sometimes referred to as “pairwise” stability. However, one may also consider group stability where blocking is allowed by arbitrary groups of agents and types. For college admissions problems with responsive preferences, pairwise stability and group stability coincide. In such environments it suffices to know the priority orderings over individual agents and the implementation of responsive DA-mechanisms is much easier than it would be for more general college preferences (e.g., substitutable preferences). From now on, we will assume that priorities are responsive, i.e., we assume that college preferences are responsive in the associated college admissions problem.8
Remark 1. Responsiveness of Priorities
Let x ∈ O, 0 < qx ≤ |N |, and x be a priority ordering. Let 2Nqx denote the set of all subsets of
N that do not exceed the capacity qx, i.e., 2Nqx ≡ {S ⊆ N : |S| ≤ qx}. Let Px denote a priority
relation on 2Nqx, i.e., Px strictly orders all sets in 2Nqx. Then, Px is responsive to x if the following
two conditions hold: (r1) for all S ∈ 2Nqx such that |S| < qx and all i ∈ N \ S, S ∪ {i} PxS and
(r2) for all S ∈ 2Nqx such that |S| < qx and all i, j ∈ N \ S, (S ∪ {i}) Px(S ∪ {j}) if and only if
i x j. When formulating (r1) we implicitly assume that each object finds all agents acceptable.
Let P( x) denote the set of all priority relations that are responsive to x.
Now, given (R, q) ∈ RN × Q and a priority relation profile PO ∈ ×x∈OP( x), an allocation a ∈ Aq is group stable under PO if there exists no coalition (consisting possibly of both agents and
types) that blocks allocation a.9
8
Correspondingly, priorities would be substitutable if college preferences are substitutable in the associated college admissions problem.
9
Formally, given (R, q) and PO, a coalition S ⊆ N ∪ O blocks a ∈ Aqif there exists an allocation b ∈ Aqsuch that
(g1) for all i ∈ S ∩ N , bi∈ S ∩ O, (g2) for all i ∈ S ∩ N , biRiai, (g3) for all x ∈ S ∩ O, i ∈ {j ∈ N : bj= x} implies
i ∈ S ∪ {j ∈ N : aj = x}, (g4) for all x ∈ S ∩ O, {j ∈ N : bj = x} Rx{j ∈ N : aj = x}, and (g5) for at least one
It is known that given (R, q) ∈ RN × Q and a responsive priority relation profile PO ∈ ×x∈OP ( x), an allocation a ∈ Aq is group stable under PO if and only if a ∈ Aq is stable
under . In other words, group stability is identical with stability for responsive priorities and the set of group stable matchings is invariant with respect to the responsive preference extensions of . This implies that for the implementation of any responsive DA-mechanism, we only need to know the priority orderings over individual agents and not the complete priority relations over sets of agents. This makes the application of responsive DA-mechanisms very easy and convenient in
real-life matching markets.
4
Simple and Intuitive Properties
A natural requirement for a mechanism is that the chosen allocation depends only on preferences over the set of available types. Given a capacity vector q, a type x is unavailable if qx= 0.
Unavailable Type Invariance: For all (R, q) ∈ RN× Q and all R0∈ RN such that R|
O+(q)∪{∅}=
R0|O+(q)∪{∅}, ϕ(R, q) = ϕ(R0, q).
By individual rationality each agent should weakly prefer his allotment to the null object (which may represent an outside option such as off-campus housing in the context of university housing allocation, or private schools or home schooling in the context of student placement in public schools).
Individual Rationality: For all (R, q) ∈ RN × Q and all i ∈ N , ϕi(R, q) Ri∅.
Next, we introduce two properties that require a mechanism to not waste any resources. First, non-wastefulness (Balinski and S¨onmez, 1999) requires that no agent prefers an available object that is not assigned to his allotment. Note that non-wastefulness corresponds to the first part (s1) of stability (for any arbitrary priority structure).
Non-Wastefulness: For all (R, q) ∈ RN × Q, all x ∈ O+(q), and all i ∈ N , if x Piϕi(R, q), then
|{j ∈ N : ϕj(R, q) = x}| = qx.
Non-wastefulness is not a weak efficiency requirement in environments where priorities can be interpreted as preferences because types (or colleges or schools) may be worse off after the departure of some agents. Or in other words, each type may prefer to have assigned as many objects of this type as possible (not caring of the identities of the students receiving this type).10 Next, we weaken non-wastefulness by requiring that no agent receives the null object while he prefers an available object that is not assigned.
Weak Non-Wastefulness: For all (R, q) ∈ RN× Q, all x ∈ O+(q), and all i ∈ N , if x Piϕi(R, q)
and ϕi(R, q) = ∅, then |{j ∈ N : ϕj(R, q) = x}| = qx.
10This may be due to avoiding the closure of a school or a college and if more positions are filled, then this may
Weak non-wastefulness is a limited efficiency requirement. For example, suppose that a central agency registers all agents who did not receive anything (or are unemployed) and all those agents report all real types (or jobs) which are acceptable to them. Then it should not be the case that some agent who receives nothing prefers one of the available real objects (or available jobs) to the null object.
Many mechanisms that are used in real life ignore agents’ preferences below their allotments (e.g., all mechanisms mentioned in Section 3). That is, an allocation does not change if an agent changes his reported preferences below his allotment. We formulate a weaker version of this invari-ance property by restricting agents’ changes below their allotments to truncations.
A truncation strategy is a preference relation that ranks the real types in the same way as the corresponding original preference relation and each real type which is acceptable under the truncation strategy is also acceptable under the original preference relation. Formally, given i ∈ N and Ri ∈ R, a strategy ¯Ri ∈ R is a truncation (strategy) of Ri if (t1) ¯Ri|O= Ri|O and (t2) A( ¯Ri) ⊆
A(Ri). Loosely speaking, a truncation strategy of Ri is obtained by moving the null object “up.”
If an agent truncates his preference relation in a way such that his allotment remains acceptable under the truncated preference relation, then truncation invariance requires that the allocation is the same under both profiles. The property is quite natural on its own in the sense that the chosen allocations do not depend on where any agent, who receives a real object, ranks the null object below his allotment.
Truncation Invariance: For all (R, q) ∈ RN × Q, all i ∈ N , and all ¯R
i ∈ Ri, if ¯Ri is a
truncation of Ri and ϕi(R, q) is acceptable under ¯Ri (i.e., ϕi(R, q) ∈ A( ¯Ri)) and ¯R = ( ¯Ri, R−i),
then ϕ( ¯R, q) = ϕ(R, q).
Our last property applies to problems when two agents compete for the same object in a maximal conflict situation, i.e., they have the same preference relation with only one acceptable real type x ∈ O and for the two problems under consideration the preference profiles are identical. Two-agent consistent conflict resolution then requires that if in these problems one of them receives the object, the conflict is resolved consistently in that it has to be the same agent in both problems who “wins the conflict” and receives the object.
We denote a preference relation with only one acceptable type x ∈ O by Rx, i.e., A(Rx) = {x}. We denote the set of all preference relations that have x ∈ O as the unique acceptable type by Rx. Two-Agent Consistent Conflict Resolution: For all R ∈ RN, all q, q0 ∈ Q, and all Rx ∈ Rx,
if Ri = Rj = Rx and {ϕi(R, q), ϕj(R, q)} = {ϕi(R, q0), ϕj(R, q0)} = {x, ∅}, then for k ∈ {i, j},
ϕk(R, q) = ϕk(R, q0).
One can interpret this property as a weak tie-breaking property because it only imposes that the tie between two agents in very specific maximal conflict situations is broken consistently in favor
of the same agent. In particular, when considering two different preference profiles R, R0 ∈ RN
where agents i and j have a maximal conflict for type x (i.e., Ri = Rj = R0i = R0j ∈ Rx and
{ϕi(R, q), ϕj(R, q)} = {ϕi(R0, q0), ϕj(R0, q0)} = {x, ∅}), two-agent consistent conflict resolution does
not imply that the same agent (out of i and j) receives object x because the assignment of x may depend on the other agents’ preferences (i.e., for k ∈ {i, j}, ϕk(R, q) 6= ϕk(R0, q0) is possible).
5
Strategy-Proofness makes the Difference
The properties we have introduced up to now are simple and intuitive. Indeed, our first result confirms this since all the real-life mechanisms we have described so far satisfy them.
Proposition 1. Let be a strict priority structure.
(i) For any priority function g, the priority mechanism based on (g, ) satisfies unavailable type invariance, individual rationality, weak non-wastefulness, two-agent consistent conflict resolution, and truncation invariance.
(ii) For any weighting function h, the linear programming mechanism based on (h, ) satisfies un-available type invariance, individual rationality, weak non-wastefulness, two-agent consistent conflict resolution, and truncation invariance.
(iii) The responsive DA-mechanism based on satisfies unavailable type invariance, individual rationality, weak non-wastefulness, two-agent consistent conflict resolution, and truncation invariance.
Note that all the properties introduced in Section 4 and used in Proposition 1 ignore the priority structure. We prove Proposition 1 in Appendix A.b
Next we characterize the class of all responsive DA-mechanisms (and not only MTB-DA-mechanisms). For our characterization of responsive DA-mechanisms we need one more property.
The well-known non-manipulability property strategy-proofness requires that no agent can ever benefit from misrepresenting his preferences.
Strategy-Proofness: For all (R, q) ∈ RN × Q, all i ∈ N , and all ¯Ri ∈ R, ϕi(R, q) Ri
ϕi(( ¯Ri, R−i), q).
It turns out that the class of mechanisms satisfying our simple and intuitive properties and strategy-proofness is very small. With the following characterization, we establish the first norma-tive foundation for the class of responsive DA-mechanisms.
Theorem 1. Responsive DA-mechanisms are the only mechanisms satisfying unavailable type invariance, individual rationality, weak non-wastefulness, two-agent consistent conflict resolution, truncation invariance, and strategy-proofness.
It is important to note that priority mechanisms and linear programming mechanisms satisfy all properties of Theorem 1 except for strategy-proofness. This demonstrates how weak all properties in Theorem 1 are, except for strategy-proofness. It is the incentive compatibility property that fa-vors responsive DA-mechanisms over all other mechanisms. We prove Theorem 1 in Appendix B.11 Appendix C contains a second characterization (Theorem 4) of responsive DA-mechanisms in a variable agents environment based on individual rationality, weak non-wastefulness, weak consis-tency, and strategy-proofness. Here weak consistency requires that agents receiving the null object should still receive the null object if some agents leave the economy with their allotments.
Remark 2. (Weak) Non-Wastefulness
Note that if a mechanism is non-wasteful, then (s1) in the definition of stability can never occur and therefore, non-wastefulness is a partial stability property. However, weak non-wastefulness, even together with individual rationality, does not imply (s1) and therefore, it is not a partial stability property. Any mechanism satisfying the properties of Theorem 1 must be stable (under a strict priority structure). The combination of the properties in Theorem 1 includes weak non-wastefulness as a partial efficiency requirement, but not as a partial stability property. However, in the above characterization (and in Theorem 4, Appendix C), weak wastefulness may be replaced by non-wastefulness (without shortening the proofs or altering the independence of the axioms). Note that similarly to responsive DA-mechanisms, priority mechanisms and linear programming mechanisms satisfy weak non-wastefulness, but in contrast to them, they do not satisfy non-wastefulness.
6
Stability makes a Difference as well
We have seen that among all the mechanisms discussed in this article, responsive DA-mechanisms are the only strategy-proof mechanisms. However, they are in fact also the only ones that respect the stability with respect to priorities. Another strategy-proof mechanism that has been discussed in school choice redesigns is the so-called top-trading cycles mechanism (Shapley and Scarf, 1974). While in most school choice redesigns (e.g., Boston and New York City) stable DA-mechanisms are implemented, only San Francisco so far has announced plans to adopt an unstable top-trading cycles based mechanism. In many applications, stability seems to be an extremely important criterium as well and it differentiates DA-mechanisms from top-trading mechanisms.
As already mentioned in our introduction, many real life applications of market design, e.g., school choice problems, have weak priority structures as inputs (see Abdulkadiro˘glu, Pathak, and Roth, 2009; Erdil and Ergin, 2008). Given a weak priority structure , stability underb imme-b diately implies non-wastefulness and individual rationality. Thus, we obtain the following first normative foundation for responsive MTB-DA-mechanisms.
Theorem 2. Let be a (weak) priority structure. Responsive MTB-DA-mechanisms are the onlyb mechanisms satisfying stability under, unavailable type invariance, two-agent consistent conflictb resolution, truncation invariance, and strategy-proofness.
A special case of a weak priority structure is the one where all agents have equal priorities for all objects (or there are no priorities given), which we denoted by ∼. Then any problem reducesb to an allocation problem with no priorities. Furthermore, it is easy to see that stability under ∼b is equivalent to non-wastefulness and individual rationality (and the properties in Theorem 2 are independent because the properties in Theorem 1 are independent).
When applying responsive DA-mechanisms to problems with weak priorities, ties have to be broken (in the special case of no priorities, all ties have to be broken). Essentially, whenever ties are present in school choice, we face trade-offs between efficiency, stability, and strategy proofness. Abdulkadiro˘glu, Pathak, and Roth (2009) show that simulations with field data from the New York City high school match and the theory favor single tie-breaking rules, i.e., indifferences are broken the same way at every school. Formally, a MTB of is a single tie-breaking (STB) if for allb x, y ∈ O and all i, j ∈ N , if i∼bxj and i∼byj, then either [i xj and i y j] or [j xi and j y i].
Now, it is interesting to ask what exactly characterizes the subclass of responsive STB-DA-mechanisms. Clearly, two-agent consistent conflict resolution in Theorem 2 adds an element of decisiveness that, together with all other properties, helps to pins down the priority structure for the underlying DA-mechanism. In order to characterize responsive STB-DA-mechanisms, the priority ordering will have to be the same for all objects and therefore we will strengthen the two-agent consistent conflict resolution property by extending the decisiveness requirement to two (possibly distinct) objects.
Strong two-agent consistent conflict resolution applies to two problems when two agents compete for two (possibly distinct) objects in maximal conflict situations, i.e., in each problem they have the same preference relation such that in one problem they find only x ∈ O acceptable and in another problem they find only y ∈ O acceptable: it then requires that if in these problems one agent receives the acceptable object and the other receives the null object, the conflict is resolved consistently in that it has to be the same agent in both problems who “wins the conflict” and receives the acceptable object.
Strong Two-Agent Consistent Conflict Resolution: For all R, R0 ∈ RN, all q, q0 ∈ Q, all
Rx∈ Rx, and all Ry ∈ Ry, if R
i = Rj = Rx, i∼bxj and {ϕi(R, q), ϕj(R, q)} = {x, ∅}, and R
0
i= R0j =
Ry, i∼byj and {ϕi(R0, q0), ϕj(R0, q0)} = {y, ∅}, then for k ∈ {i, j}, ϕk(R, q) = ∅ ⇔ ϕk(R0, q0) = ∅.
Now a straightforward consequence of Theorem 2 is the following normative foundation for responsive STB-DA-mechanisms.
Theorem 3. Let be a (weak) priority structure. Responsive STB-DA-mechanisms are the onlyb mechanisms satisfying stability under , unavailable type invariance, strong two-agent consistentb conflict resolution, truncation invariance, and strategy-proofness.
7
Conclusion
Our model is that of assigning types (school seats) to a set of agents (students). Agents have strict preferences over types and whenever we consider agents’ strict priorities for a certain type, we capture them by an ordering of all agents. Of course, since sets of agents are assigned to an type, in general priorities may depend on the whole set of agents and not only on the ordering of individual agents. However, if priorities over sets of agents are responsive with respect to the strict priority ordering over individual agents, then in determining stable allocations we only need to know the priority orderings over individual agents. It is exactly this responsiveness-to-individual-priorities feature that makes the agent-proposing deferred-acceptance mechanism easily applicable in practice (Roth, 2008).
Kojima and Manea (2010) provide two characterizations of deferred acceptance mechanisms with so-called acceptant substitutable priorities (a larger class of mechanisms than the class of responsive DA-mechanisms which is based on priorities that are determined by a choice function that reflects substitutability in priorities over sets of agents; see also Hatfield and Milgrom, 2005). For this class of DA-mechanisms, priority orderings over individual agents do not suffice—the priorities over sets of agents must be known. Furthermore, their characterizations use two new monotonicity properties which are violated by priority mechanisms and linear programming mech-anisms. Indeed, it can be seen that those classes of mechanisms violate all their properties in their first characterization and two of the three properties in their second characterization. While the class of DA-mechanisms with acceptant substitutable priorities is sometimes employed in real-life markets, many applications involving DA-mechanisms are based on responsive priorities or slight generalizations thereof. Thus, providing a rationale for this class of mechanisms is important. We characterized responsive DA-mechanisms using very simple and intuitive properties plus strategy-proofness.
References
Abdulkadiro˘glu, A., Pathak, P., and A.E. Roth (2009): Strategy-Proofness versus Efficiency in Matching with Indifferences: Redesigning the NYC High School Match, American Economic Review 99, 1954–1978.
Abdulkadiro˘glu, A., Pathak, P., Roth, A.E., and T. S¨onmez (2006): Changing the Boston School Choice Mechanism: Strategy-Proofness as Equal Access, Working Paper, Harvard University.
Abdulkadiro˘glu, A., T. S¨onmez (2003): School Choice: a Mechanism Design Approach, American Economic Review 93, 729–747.
Balinski, M., and T. S¨onmez (1999): A Tale of Two Mechanisms: Student Placement, Journal of Economic Theory 84, 73–94.
Dubins, L.E., and D.A. Freedman (1981): Machiavelli and the Gale-Shapley Algorithm, American Mathematical Monthly 88, 485–494.
Ehlers, L. (2002): Coalitional Strategy-Proof House Allocation, Journal of Economic Theory 105, 298–317.
Ehlers, L. (2008): Truncation Strategies in Matching Markets, Mathematics of Operations Re-search 33, 327–335.
Ehlers, L., and A. Erdil (2010): Efficient Assignment Respecting Priorities, Journal of Economic Theory 145, 1269–1282.
Ehlers, L., and B. Klaus (2006): Efficient Priority Rules, Games and Economic Behavior 55, 372–384.
Ehlers, L., and B. Klaus (2007): Consistent House Allocation, Economic Theory 30, 561–574. Ehlers, L., and B. Klaus (2009): Allocation via Deferred Acceptance, CIREC Cahier 17-2009. Erdil, A., and H. Ergin (2008): What’s the Matter with Tie-Breaking? Improving Efficiency in
School Choice, American Economic Review 98, 669–689.
Ergin, H.˙I. (2002): Efficient Resource Allocation on the Basis of Priorities, Econometrica 70, 2489–2497.
Ergin, H.˙I., and T. S¨onmez (2006): Games of School Choice under the Boston Mechanism, Journal of Public Economics 90, 215–237.
Gale, D., and L. Shapley (1962): College Admissions and the Stability of Marriage, American Mathematical Monthly 69, 9–15.
Hatfield, J.W., and P. Milgrom (2005): Matching with Contracts, American Economic Review 95, 913–935.
Kesten, O. (2009): Coalitional Strategy-Proofness and Resource Monotonicity for House Alloca-tion Problems, InternaAlloca-tional Journal of Game Theory 38, 17–21.
Kojima, F., and M. Manea (2010): Axioms for Deferred Acceptance, Econometrica 78, 633–653. P´apai, S. (2000): Strategyproof Assignment by Hierarchical Exchange, Econometrica 68, 1403–
1433.
Pathak, P., and T. S¨onmez (2008): Leveling the Playing Field: Sincere and Sophisticated Players in the Boston Mechanism, American Economic Review 98, 1636–52.
Roth, A.E. (1982): The Economics of Matching: Stability and Incentives, Mathematics of Opera-tions Research 7, 617–628.
Roth, A.E. (1984a): The Evolution of the Labor Market for Medical Interns and Residents: A Case Study in Game Theory, Journal of Political Economy 92, 991–1016.
Roth, A.E. (1984b): Misrepresentation and Stability in the Marriage Problem, Journal of Eco-nomic Theory 34, 383–387.
Roth, A.E. (1991): A Natural Experiment in the Organization of Entry Level Labor Markets: Regional Markets for New Physicians and Surgeons in the U.K., American Economic Review 81, 415–440.
Roth, A.E. (2008): Deferred Acceptance Algorithm: History, Theory, Practice, and Open Ques-tions, International Journal of Game Theory 36, 537–569.
Roth, A.E., and M.A.O. Sotomayor (1990): Two-Sided Matching: A Study in Game-Theoretic Modeling and Analysis. Econometric Society Monograph Series. New York: Cambridge University Press.
Shapley, L., and H. Scarf (1974): On Cores and Indivisibility, Journal of Mathematical Economics 1, 23–37.
S¨onmez, T., and M.U. ¨Unver (2011): Matching, Allocation, and Exchange of Discrete Resources, Jess Benhabib, Alberto Bisin, and Matthew Jackson (eds.) Handbook of Social Economics, Elsevier.
Thomson, W. (2009): Consistent Allocation Rules, book manuscript.
Appendix
A
Proof of Proposition 1
Let be a strict priority structure.
In showing (i), let g be a priority function and consider the priority mechanism M(g, ) based
on g and . By definition, M(g, ) satisfies unavailable type invariance and individual rationality. For weak non-wastefulness, suppose that for (R, q), x ∈ O+(q) and i ∈ N we have
x Pi Mi(g, )(R, q) and Mi(g, )(R, q) = ∅. If |{j ∈ N : Mj(g, )(R, q) = x}| < qx, then (i, x) is
an individually rational (a, b)-match and the priority mechanism M(g, ) could not have stopped without matching i and x.
For two-agent consistent conflict resolution, let (R, q) and (R, q0) be such that Ri = Rj = Rx
and {Mi(g, )(R, q), Mj(g, )(R, q)} = {Mi(g, )(R, q0), Mj(g, )(R, q0)} = {x, ∅}. Since rank(x, Ri) =
1 = rank(x, Rj), (i, x) is an (1, b)-match and (j, x) is an (1, b0)-match where b = rank(i, x) and
b0 = rank(j, x). Because x is strict, b 6= b0 and we obtain from {Mi(g, )(R, q), M (g, )
j (R, q)} =
{Mi(g, )(R, q0), Mj(g, )(R, q0)} = {x, ∅} both Mi(g, )(R, q) = Mi(g, )(R, q0) and Mj(g, )(R, q) = Mj(g, )(R, q0), the desired conclusion.
For truncation invariance, let (R, q), i ∈ N , and ¯Ri be such that Mi(g, )(R, q) ∈ A( ¯Ri) and
¯
( ¯Ri, R−i, q) the priority mechanism M(g, ) matches identically all agent-type pairs and we obtain
M(g, )(R, q) = M(g, )( ¯Ri, R−i, q).
In showing (ii), note that all properties can be easily formulated for random mechanisms (for instance by putting positive probability only on individually rational allocations). Let h be a weighting function and consider the linear programming mechanism M(h, ) based on h and . By definition, M(h, ) satisfies unavailable type invariance and individual rationality.
For weak non-wastefulness, suppose that for (R, q), x ∈ O+(q) and i ∈ N we have
x PiMi(h, )(R, q) and M (h, )
i (R, q) = ∅. If |{j ∈ N : M (h, )
j (R, q) = x}| < qx, then (i, x) is some
individually rational (a, b)-match with h(a, b) > 0. Now M(h, )(q, R) could not have maximized the score among all individually rational allocations.
For two-agent consistent conflict resolution, let (R, q) and (R, q0) be such that Ri = Rj = Rx
and {Mi(h, )(R, q), Mj(h, )(R, q)} = {Mi(h, )(R, q0), Mj(h, )(R, q0)} = {x, ∅}. Since rank(x, Ri) =
1 = rank(x, Rj), (i, x) is an (1, b)-match and (j, x) is an (1, b0)-match where b = rank(i, x)
and b0 = rank(j, x). Because x is strict, b 6= b0 and h(1, b) 6= h(1, b0). Thus, we obtain
from {Mi(h, )(R, q), Mj(h, )(R, q)} = {Mi(h, )(R, q0), Mj(h, )(R, q0)} = {x, ∅} both Mi(h, )(R, q) = Mi(h, )(R, q0) and Mj(h, )(R, q) = Mj(h, )(R, q0), the desired conclusion.
For truncation invariance, let (R, q), i ∈ N , and ¯Ri be such that Mi(h, )(R, q) ∈ A( ¯Ri) and ¯Ri
is a truncation of Ri. Note that this implies Mi(h, )(R, q) 6= ∅. Hence, under (R, q) and ( ¯Ri, R−i, q)
the allocation M(h, )(R, q) maximizes the score among all individually rational allocations and we obtain M(h, )(R, q) = M(h, )( ¯Ri, R−i, q) (we refer to Ehlers (2008) for random mechanisms
satisfying truncation invariance).
In showing (iii), consider the responsive DA-mechanism based on . It is easy to see that DA satisfies unavailable type invariance, individual rationality, weak non-wastefulness, two-agent consistent conflict resolution, and truncation invariance.
B
Proof of Theorem 1
It is easy to verify that responsive DA-mechanisms satisfy the properties of Theorem 1. Conversely, let ϕ be a mechanism satisfying the properties of Theorem 1. First, we “calibrate/construct the priority structure using maximal conflict preference profiles.”
Let x ∈ O and let Rx ∈ Rx (i.e., A(Rx) = {x}). Let R∅ ∈ R be such that A(R∅) = ∅.
For any S ⊆ N , let Rx
S = (Rxi)i∈S such that for all i ∈ S, Rxi = Rx, and similarly R∅S= (R ∅ i)i∈S
such that for all i ∈ S, R∅i = R∅.
Let 1xdenote the capacity vector q such that qx= 1 and for all z ∈ O\{x}, qz = 0. Similarly, for
y ∈ O \ {x} let 1xy denote the capacity vector q such that qx= 1, qy = 1, and for all z ∈ O \ {x, y},
Consider the problem (RxN, 1x). By weak non-wastefulness, for some j ∈ N , ϕj(RxN, 1x) = x,
say j = 1. Then, for all i ∈ N \ {1}, we set 1 x i.
Next consider the problem ((R1∅, Rx
−1), 1x). By weak non-wastefulness and individual rationality,
for some j ∈ N \ {1}, ϕj((R∅1, Rx−1), 1x) = x, say j = 2. Then, for all i ∈ N \ {1, 2}, we set 2 x i.
By induction, we obtain x for any type x and thus a priority structure = ( x)x∈O.
Lemma 1. For all R ∈ RN and all x ∈ O, if for some j ∈ N , ϕj(R, 1x) = x, then for all
i ∈ N \ {j}, x ∈ A(Ri) implies j xi.
Proof. Let R ∈ RN and x ∈ O. Without loss of generality, suppose 1 x 2 x · · · x n. Let
S = {i ∈ N : x ∈ A(Ri)} and let j = min S. We prove Lemma 1 by showing that ϕj(R, 1x) = x. In
the sequel, when using two-agent consistent conflict resolution we often also implicitly apply weak non-wastefulness and individual rationality.
Note that for all i ∈ N \S, ∅Pix. We partition the set N \S into the “lower” set L = {1, . . . , j−1}
(possibly L = ∅) and the “upper” set U = N \ (L ∪ S) (possibly U = ∅). Note that by unavailable type invariance, ϕ(R, 1x) = ϕ((R∅L, RxS, R∅U), 1x).
By the construction of x, ϕj((R∅L, RxS∪U), 1x) = x. Hence, if U = ∅, then ϕj(R, 1x) = x and
for all i ∈ N , x ∈ A(Ri) implies j x i.
Step 1 : Let k ∈ U . We prove that ϕj((R∅L, RxS∪(U \{k}), R∅k), 1x) = ϕj((R∅L, RxS∪U), 1x) = x.
Let y ∈ O \ {x}. By two-agent consistent conflict resolution, ϕj((RL∅, RxS∪U), 1xy) = x. Let
Rk0 ∈ R be such that R0
k : y x ∅ . . .. By strategy-proofness, ϕk((R∅L, RxS∪(U \{k}), R 0
k), 1xy) 6=
x. If S ∪ (U \ {k}) = {j}, then by weak non-wastefulness and individual rational-ity, ϕj((R∅L, RxS∪(U \{k}), R0k), 1xy) = x. Otherwise, by two-agent consistent conflict
resolu-tion, ϕj((R∅L, RxS∪(U \{k}), R 0
k), 1xy) = x. By weak non-wastefulness and individual rationality,
ϕk((R∅L, RxS∪(U \{k}), R 0
k), 1xy) = y.
Let R00k ∈ R be such that R00k : y ∅ x . . . and R00k|O = R0k|O. By strategy-proofness,
ϕk((R∅L, RxS∪(U \{k}), R 00
k), 1xy) = y. By weak non-wastefulness and individual rationality, for some
l ∈ S ∪ (U \ {k}), ϕl((RL∅, RxS∪(U \{k}), R 00
k), 1xy) = x.
Now R00k is a truncation of R0k and both y ∈ A(R00k) and ϕk((RL∅, RxS∪(U \{k}), R0k), 1xy) =
y. Thus, ϕ((R∅L, RxS∪(U \{k}), Rk00), 1xy) = ϕ((R∅L, RxS∪(U \{k}), R 0
k), 1xy). Hence, j = l and
ϕj((RL∅, RxS∪(U \{k}), R00k), 1xy) = x.
By individual rationality, ϕk((R∅L, RxS∪(U \{k}), Rk00), 1x) 6= x. If S ∪ (U \ {k}) = {j}, then by
weak non-wastefulness and individual rationality, ϕj((R∅L, RxS∪(U \{k}), R00k), 1x) = x. Otherwise, by
two-agent consistent conflict resolution, ϕj((R∅L, RxS∪(U \{k}), R00k), 1x) = x. Thus, by unavailable
type invariance, ϕj((R∅L, RxS∪(U \{k}), R∅k), 1x) = ϕj((R∅L, RxS∪(U \{k}), R00k), 1x) = x.
Steps 2, . . . l: Let U = {k1, . . . , kl}. Then using the same arguments as above, it follows that
x = ϕj((R∅L, RxS∪U), 1x) = ϕj((R∅L, RxS∪(U \{k1}), R ∅ k1), 1x) = ϕj((R ∅ L, RxS∪(U \{k1,k2}), R ∅ {k1,k2}), 1x) =
. . . = ϕj((R∅L, RxS∪{kl}, R ∅ U \{kl}), 1x) = ϕj((R ∅ L, RxS, R ∅
U), 1x) = ϕj(R, 1x). Hence, we obtain the
desired result that ϕj(R, 1x) = x.
For any number qx ∈ {0, 1, . . . , qx}, let qx◦ 1x = (qx, 0O\{x}) denote the capacity vector with
exactly qx copies of object x. The following lemma is the extension of Lemma 1 to the general
capacity case.
Lemma 2. For all R ∈ RN and all qx∈ {1, . . . , qx}, if T = {j ∈ N : ϕj(R, qx◦ 1x) = x}, then for
all j ∈ T and all i ∈ N \ T , x ∈ A(Ri) implies j x i.
Proof. Let R ∈ RN, qx ∈ {1, . . . , qx}, T = {j ∈ N : ϕj(R, qx◦ 1x) = x}, and S = {i ∈ N :
x ∈ A(Ri)}. By individual rationality, T ⊆ S. Let j ∈ T and i ∈ S \ T . We prove Lemma 2 by
showing that j x i. In the sequel, when using two-agent consistent conflict resolution we often
also implicitly apply weak non-wastefulness and individual rationality. If qx = 1, then by Lemma 1, j xi.
Next, suppose that qx = 2. Note that by unavailable type invariance, ϕ(R, qx ◦ 1x) =
ϕ((RxS, R∅−S), qx◦ 1x). Let k ∈ S be such that ϕk((RxS, R ∅
−S), 1x) = x. By Lemma 1, k x i.
By two-agent consistent conflict resolution, ϕk((RxS, R∅−S), qx◦ 1x) = x. Hence, k ∈ T and k xi.
By weak non-wastefulness and individual rationality, there exists h ∈ T \ {k} such that ϕh((RxS, R
∅
−S), qx ◦ 1x) = x. Let y ∈ O \ {x} and R0k ∈ R be such that R0k : y x ∅ . . ..
By unavailable type invariance, ϕ((RS\{k}x , R∅−S, R0k), qx ◦ 1x) = ϕ((RxS, R∅−S), qx ◦ 1x). Hence,
ϕh((RxS\{k}, R∅−S, Rk0), qx◦ 1x) = x.
Next, we replace one object of type x by one object of type y. By weak non-wastefulness and individual rationality, ϕk((RxS\{k}, R∅−S, R0k), 1xy) = y. By two-agent consistent conflict resolution,
ϕh((RxS\{k}, R∅−S, R 0
k), 1xy) = x. By truncation invariance, ϕh((RxS\{k}, R−S∅ , R y
k), 1xy) = x. By
two-agent consistent conflict resolution, ϕh((RxS\{k}, R∅−S, R y
k), 1x) = x. Hence, by Lemma 1, h xi.
Now by induction on qx the conclusion of Lemma 2 follows.
Lemma 3. For all (R, q) ∈ RN × Q, ϕ(R, q) is stable under .
Proof. Let (R, q) ∈ RN × Q and assume that ϕ(R, q) is not stable under . Then, there exists an agent-object pair (i, x) ∈ N × O ∪ {∅} such that x Piϕi(R, q) and (s1) |{j ∈ N : ϕj(R, q) = x}| < qx
or (s2) there exists k ∈ N such that ϕk(R, q) = x and i xk. By individual rationality, x 6= ∅.
Let ¯R ∈ RN be such that (a) for all j ∈ N such that ϕ
j(R, q) 6= ∅, ¯Rj is a truncation of Rj
such that there exists no y ∈ O \ {ϕj(R, q)} with ϕj(R, q) ¯Rj y ¯Rj ∅ and (b) for all j ∈ N such
that ϕj(R, q) = ∅, ¯Rj = Rj. (By individual rationality, ¯Rj in (a) is well-defined as truncation of
Rj.) By truncation invariance, ϕ( ¯R, q) = ϕ(R, q) and (i, x) ∈ N × O is such that x ¯Piϕi( ¯R, q) and
Let S = {j ∈ N : x ¯Pj ϕj( ¯R, q)}. Note that i ∈ S. Let T = {j ∈ N : ϕj( ¯R, q) = x}. By
unavailable type invariance, ϕ( ¯R, qx◦ 1x) = ϕ(( ¯RxS∪T, ¯R∅N \(S∪T )), qx◦ 1x).
Step 1 : Assume that for j ∈ S, ϕj( ¯R, qx ◦ 1x) = x. Then, by strategy-proofness and
in-dividual rationality, ϕj(( ¯R−j, Rjx), qx ◦ 1x) = x and ϕj(( ¯R−j, Rxj), q) = ∅. By weak
non-wastefulness, |{k ∈ N : ϕk(( ¯R−j, Rxj), q) = x}| = qx. By individual rationality there
exists l ∈ N such that x ∈ A( ¯Rl), and ϕl(( ¯R−j, Rjx), q) = x and ϕl(( ¯R−j, Rxj), qx ◦
1x) = ∅. By Lemma 2, j x l. Next, by strategy-proofness, ϕl(( ¯R−{j,l}, Rjx, Rxl), q) =
x. By unavailable type invariance, ϕ(( ¯R−{j,l}, Rjx, Rxl), qx ◦ 1x) = ϕ(( ¯R−j, Rxj), qx ◦ 1x);
in particular, ϕj(( ¯R−{j,l}, Rxj, Rxl), qx ◦ 1x) = x and ϕl(( ¯R−{j,l}, Rjx, Rxl), qx ◦ 1x) = ∅.
Thus, ϕj(( ¯R−{j,l}, Rxj, Rxl), q) = ∅ would contradict two-agent consistent conflict
resolu-tion because then {ϕj(( ¯R−{j,l}, Rxj, Rlx), q), ϕl(( ¯R−{j,l}, Rjx, Rlx), q)} = {ϕj(( ¯R−{j,l}, Rxj, Rxl), qx ◦
1x), ϕl(( ¯R−{j,l}, Rxj, Rxl), qx◦ 1x)} = {x, ∅}. Hence, ϕj(( ¯R−{j,l}, Rxj, Rxl), q) = x. Thus, there
ex-ists an agent j2 ∈ N \ {j, l} such that ϕj2(( ¯R−{j,l}, R
x j, Rxl), q) = ∅ and ϕj2(( ¯R−{j,l}, R x j, Rxl), qx◦ 1x) = x. However, R¯j2 = R x
j2 would contradict two-agent consistent conflict
resolu-tion because then {ϕj2(( ¯R−{j,l}, R
x j, Rxl), q), ϕl(( ¯R−{j,l}, Rjx, Rxl), q)} = {ϕj2(( ¯R−{j,l}, R x j, Rxl), qx ◦ 1x), ϕl(( ¯R−{j,l}, Rxj, Rxl), qx◦ 1x)} = {x, ∅}. Hence, ¯Rj2 6= R x j2.
Steps 2, . . . : Step 2 replicates Step 1 with the starting preference profile ( ¯R−{j,l}, Rxj, Rxl) and with
agent j2 in the role of agent j. Throughout the step, we strictly increase the number of agents with
preferences Rx (at least by agent j2). Furthermore, the step ends with the existence of another
agent j3 with ¯Rj3 6= R
x
j3 with whom we proceed to Step 3. Since the number of agents is finite, we
obtain a contradiction in finitely many steps.
Final Step: With the previous steps we have established that for all j ∈ S, ϕj( ¯R, qx◦ 1x) = ∅. Note
that by construction of ¯R, for all j ∈ N \ (S ∪ T ), ∅ ¯Pjx. Hence, by individual rationality, for all
j ∈ N \ (S ∪ T ), ϕj( ¯R, qx◦ 1x) = ∅. Note that by individual rationality, for all j ∈ T , x ∈ A( ¯Rj).
Hence, by weak non-wastefulness, for all j ∈ T , ϕj( ¯R, qx◦ 1x) = x and |T | = qx. By Lemma 2, we
have that (∗) for all j ∈ T and all l ∈ S, j xl.
Recall that T = {j ∈ N : ϕj( ¯R, q) = x}. Hence, (s1’) cannot be the case. Thus, by (s2’) there
exists k ∈ N such that ϕk( ¯R, q) = x and i x k. Recall that i ∈ S and that ϕk( ¯R, q) = x implies
that k ∈ T so that i x k contradicts (∗).
So far we have established that for any mechanism ϕ that satisfies the properties of Theorem 1, there exists a priority ordering such that for any (R, q) ∈ RN × Q, ϕ(R, q) is stable under .
Hence, in the terminology of two-sided matching, the mechanism ϕ picks a stable allocation for the many-to-one two-sided market where types have responsive preferences over sets of agents who consume the objects based on the priority structure and agents have strict preferences over objects based on preferences R (see Roth and Sotomayor, 1990, Chapter 5). For these markets it is well-known that the responsive DA-mechanism is the only strategy-proof stable matching mechanism.
For completeness, we provide a proof which uses some standard results from many-to-one two-sided markets with responsive preferences.
Lemma 4. For all (R, q) ∈ RN × Q, ϕ(R, q) = DA (R, q).
Proof. Let (R, q) ∈ RN× Q and assume that ϕ(R, q) 6= DA (R, q). By Lemma 3, ϕ(R, q) is stable
under . Thus, since DA (R, q) is the agent-optimal stable matching (Roth and Sotomayor, 1990, Corollary 5.9), for all i ∈ N , DA i (R, q) Riϕi(R, q). Since ϕ(R, q) 6= DA (R, q), there exists j ∈ N
such that DA j(R, q) Pjϕj(R, q). By individual rationality, ϕj(R, q) Rj ∅. Thus, DA j (R, q) 6= ∅.
Let ¯Rj be such that ¯Rj|O= Rj|O and there is no x ∈ O \ {DA j(R, q)} such that DA j (R, q) ¯Rj
x ¯Rj∅. Since ¯Rj is a truncation of Rj and DA satisfies truncation invariance, DA (( ¯Rj, R−j), q) =
DA (R, q). Thus, DA j (( ¯Rj, R−j), q) 6= ∅.
By Lemma 3, ϕj(( ¯Rj, R−j), q) is stable under . By Roth and Sotomayor (1990, Corollary 5.9),
DA j (( ¯Rj, R−j), q)Rjϕj(( ¯Rj, R−j), q). By DA j (( ¯Rj, R−j), q) 6= ∅ and the fact that the set of agents
receiving the null object is identical for any two stable matchings (Roth and Sotomayor, 1990, Theorem 5.12), we have ϕj(( ¯Rj, R−j), q) 6= ∅. Now, by the definition of ¯Rj, ϕj(( ¯Rj, R−j), q) =
DA j (( ¯Rj, R−j), q). Hence, ϕj(( ¯Rj, R−j), q) = DA j (( ¯Rj, R−j), q) = DA j(R, q) Pj ϕj(R, q), which
contradicts strategy-proofness.
C
Variable Agents and Weak Consistency
Let N again denote the finite set of agents, but the set of agents who are present in a problem can vary. We define the set of all nonempty subsets of N by N ≡ {M : M ⊆ N and M 6= ∅}. An (allocation) problem (with capacity constraints) now consists of a set N0∈ N of agents, a preference profile R ∈ RN0, and a capacity vector q. We denote the set of all problems byS
N0∈NRN 0
×Q. We adjust our previous model, definitions, and properties by simply replacing the domain of problems RN× Q by the variable population domainS
N0∈NRN 0
× Q. Given N0 ∈ N and R ∈ RN0
, for any M0∈ N such that M0 ⊆ N0, let RM0 denote the profile (Ri)i∈M0.
A requirement for a mechanism that is very much in the spirit of unavailable type invariance is unassigned type invariance: the chosen allocation does not depend on the unconsumed or unas-signed objects. Given a problem (R, q) ∈ S
N0∈NRN 0
× Q, we define by q(ϕ(R, q)) the capacity vector of assigned objects: for all x ∈ O, qx(ϕ(R, q)) = |{j ∈ N0: ϕj(R, q) = x}|.
Unassigned Type Invariance: For all (R, q) ∈S
N0∈N RN 0
× Q, ϕ(R, q) = ϕ(R, q(ϕ(R, q))). Consistency is one of the key properties in many frameworks with variable population scenarios. Thomson (2009) provides an extensive survey of consistency in various applications. Consistency requires that if some agents leave a problem with their allotments, then the mechanism should allocate the remaining objects among the agents who did not leave in the same way as in the original problem.
Consistency: For all M0, N0 ∈ N such that M0 ⊆ N0, all R ∈ RN0
, all q ∈ Q, and all i ∈ M0, ϕi(R, q) = ϕi(RM0, ˜q) where ˜qx = qx− |{j ∈ N0\ M0 : ϕj(R, q) = x}| for all x ∈ O.
It follows from Ergin (2002, Theorem 1)—for a fixed capacity vector q—that the only responsive DA-mechanisms satisfying consistency are the ones with an “acyclic” priority structure.12
The following property is a weak consistency property that all responsive DA-mechanisms sat-isfy. It requires that if some agents leave a problem with their allotments, then an agent who did not leave and who received the null object, still receives the null object. In other words, allocations only need to be consistent with respect to the agents who receive the null object.
Weak Consistency: For all M0, N0 ∈ N such that M0 ⊆ N0, all R ∈ RN0
, all q ∈ Q, and all i ∈ M0, if ϕi(R, q) = ∅, then ϕi(RM0, ˜q) = ∅ where ˜qx= qx− |{j ∈ N0\ M0 : ϕj(R, q) = x}| for all
x ∈ O.
Theorem 4. Responsive DA-mechanisms are the only mechanisms satisfying unassigned type in-variance, individual rationality, weak non-wastefulness, weak consistency, and strategy-proofness. Below we adjust the definition of priority mechanisms and linear programming mechanisms to the variable population framework such that they satisfy all properties in Theorem 4 except for strategy-proofness (like Proposition 1 for Theorem 1).
For a strict priority structure and a priority function g, the priority mechanism based on (g, ) matches for any problem (R, q) ∈S
N0∈NRN 0
× Q all agent-type pairs that have the highest priority according to g subject to individual rationality and the quotas. Note that the ranking of unavailable objects is maintained in the agents’ preferences and the ranking of agents belonging to N \N0 is maintained in .
Similarly, for a strict priority structure and a weighting function h, the linear programming mechanism based on (h, ) chooses for any problem (R, q) ∈ S
N0∈NRN 0
× Q an individually rational allocation with maximal score among all individually rational allocations. Note that here again the ranking of unavailable objects is maintained in the agents’ preferences and the ranking of agents belonging to N \N0 is maintained in .
Analogously to Theorem 2 we obtain a foundation of responsive MTB-DA-mechanisms using the previous characterization (Theorem 4).
Theorem 5. Let be a (weak) priority structure. Responsive MTB-DA-mechanisms are theb only mechanisms satisfying stability under , unassigned type invariance, weak consistency, andb strategy-proofness.
12
Proof of Theorem 4
It is easy to verify that responsive DA-mechanisms satisfy the properties of Theorem 4. Conversely, let ϕ be a mechanism satisfying these properties. First, we “calibrate/construct the priority struc-ture using maximal conflict preference profiles.”
Let x ∈ O and let Rx ∈ Rx (i.e., A(Rx) = {x}). Let R∅ ∈ R be such that A(R∅) = ∅.
For any S ⊆ N , let RxS = (Rxi)i∈S such that for all i ∈ S, Rix = Rx, and similarly R∅S= (Ri∅)i∈S
such that for all i ∈ S, R∅i = R∅.
Let 1x denote the capacity vector q such that qx= 1 and for all z ∈ O \ {x}, qz= 0.
Consider the problem (RxN, 1x). By weak non-wastefulness, for some j ∈ N , ϕj(RxN, 1x) = x,
say j = 1. Then, for all i ∈ N \ {1}, we set 1 x i.
Next consider the problem (RxN \{1}, 1x). By weak non-wastefulness and individual rationality,
for some j ∈ N \ {1}, ϕj(RxN \{1}, 1x) = x, say j = 2. Then, for all i ∈ N \ {1, 2}, we set 2 xi.
By induction, we obtain x for any type x and thus a priority structure = ( x)x∈O.
Lemma 5. For all N0 ∈ N , all R0 ∈ RN0, and all x ∈ O, if for some j ∈ N , ϕ
j(R0, 1x) = x, then
for all i ∈ N0\ {j}, x ∈ A(R0
i) implies j x i.
Proof. Let N0 ∈ N , R0 ∈ RN0
, and x ∈ O. Without loss of generality, suppose 1 x 2 x · · · x n.
Let S = {i ∈ N0: x ∈ A(R0i)} and let j = min S. We prove Lemma 5 by showing that ϕj(R0, 1x) =
x.
By weak non-wastefulness and individual rationality, for some l ∈ S, ϕl(R0, 1x) = x. Assume, for
a contradiction, that l 6= j. Thus, ϕj(R0, 1x) = ∅. Hence, by weak consistency, ϕj(R0{j,l}, 1x) = ∅.
By weak non-wastefulness, ϕl(R0{j,l}, 1x) = x. By strategy-proofness, ϕl((R0j, Rxl), 1x) = x and
ϕj((Rjx, Rxl), 1x) = ∅. By weak non-wastefulness, ϕl((Rxj, Rxl), 1x) = x.
Let L = {1, . . . , j − 1} (possibly L = ∅). By the construction of x, ϕj(RxN \L, 1x) = x. Thus,
ϕl(RxN \L, 1x) = ∅. Hence, by weak consistency, ϕl(Rx{j,l}, 1x) = ∅. By weak non-wastefulness,
ϕj(Rx{j,l}, 1x) = x. Since Rx{j,l} = (Rjx, Rxl) and l 6= j, we have established a contradiction. Thus,
l = j and ϕj(R0, 1x) = x.
Let ˆR = {Ri ∈ R : |A(Ri)| ≤ 1}.
Lemma 6.
(a) For all N0∈ N and all (R, q) ∈ RN0 × Q, if |N0| = 2, then ϕ(R, q) is stable under .
(b) For all N0 ∈ N and all (R, q) ∈ ˆRN0
× Q, ϕ(R, q) is stable under .
Proof. In order to show (a), suppose that ϕ(R, q) is not stable under . Then, there exists an agent-object pair (i, x) ∈ N0×O ∪{∅} such that xPiϕi(R, q) and (s1) |{j ∈ N0: ϕj(R, q) = x}| < qx
